Aplicação de machine learning em dados climáticos para a predição de incêndios na caatinga piauiense

Abstract

Wildfires are a major threat to the Caatinga biome, causing significant environmental, social, and economic impacts. In this context, this study aimed to develop a machine learning-based predictive model to estimate the daily occurrence of wildfires in the state of Piauí using meteorological data. Historical wildfire records and weather station data were used, including variables such as air temperature, relative humidity, precipitation, atmospheric pressure, and wind speed. After data preprocessing and integration, the Light Gradient Boosting Machine algorithm was applied using training, validation, and testing datasets, together with decision threshold optimization to improve the model's predictive performance. The results showed an accuracy of 82%, precision of 79%, recall of 88%, an F1-score of 83%, an area under the curve of 90%, and a Cohen's kappa coefficient of 64%, demonstrating the model's ability to identify wildfire events. As a practical application, the model was used to generate wildfire risk maps, highlighting its potential as a decision-support tool for environmental monitoring and wildfire prevention.